[Misc] reuse num_tokens_across_dp of get_dp_padding to avoid unnecessary dp all reduce in set_forward_context (#18935)

Signed-off-by: Tyler Michael Smith <tysmith@redhat.com>
Co-authored-by: zhuhaoran <zhuhaoran.zhr@alibaba-inc.com>
Co-authored-by: Tyler Michael Smith <tysmith@redhat.com>
This commit is contained in:
zhrrr
2025-06-01 19:41:18 +00:00
committed by GitHub
co-authored by zhuhaoran Tyler Michael Smith
parent 432ec9926e
commit d6fd3a33b8
2 changed files with 47 additions and 18 deletions
+19 -8
View File
@@ -47,8 +47,12 @@ class DPMetadata:
return num_tokens_tensor
@staticmethod
def make(parallel_config: ParallelConfig, attn_metadata: Any,
num_tokens: int) -> "DPMetadata":
def make(
parallel_config: ParallelConfig,
attn_metadata: Any,
num_tokens: int,
num_tokens_across_dp: Optional[torch.Tensor] = None
) -> "DPMetadata":
assert parallel_config.data_parallel_size > 1
dp_size = parallel_config.data_parallel_size
@@ -62,10 +66,15 @@ class DPMetadata:
# for v1 attention backends or no attn_metadata
batchsize = num_tokens
num_tokens_tensor = DPMetadata.num_tokens_across_dp(
batchsize, dp_size, dp_rank)
max_tokens_across_dp_cpu = torch.max(num_tokens_tensor)
cu_tokens_across_dp_cpu = torch.cumsum(num_tokens_tensor, dim=0)
# If num_tokens_across_dp is None, it will be computed by all_reduce
# Otherwise, num_tokens_across_dp[dp_rank] should be equal to batchsize
assert (num_tokens_across_dp is None
or num_tokens_across_dp[dp_rank] == batchsize)
if num_tokens_across_dp is None:
num_tokens_across_dp = DPMetadata.num_tokens_across_dp(
batchsize, dp_size, dp_rank)
max_tokens_across_dp_cpu = torch.max(num_tokens_across_dp)
cu_tokens_across_dp_cpu = torch.cumsum(num_tokens_across_dp, dim=0)
return DPMetadata(max_tokens_across_dp_cpu, cu_tokens_across_dp_cpu)
@@ -101,7 +110,8 @@ def get_forward_context() -> ForwardContext:
def set_forward_context(attn_metadata: Any,
vllm_config: VllmConfig,
virtual_engine: int = 0,
num_tokens: Optional[int] = None):
num_tokens: Optional[int] = None,
num_tokens_across_dp: Optional[torch.Tensor] = None):
"""A context manager that stores the current forward context,
can be attention metadata, etc.
Here we can inject common logic for every model forward pass.
@@ -114,7 +124,8 @@ def set_forward_context(attn_metadata: Any,
if vllm_config.parallel_config.data_parallel_size > 1 and (
attn_metadata is not None or num_tokens is not None):
dp_metadata = DPMetadata.make(vllm_config.parallel_config,
attn_metadata, num_tokens or 0)
attn_metadata, num_tokens or 0,
num_tokens_across_dp)
global _forward_context
prev_context = _forward_context